Publication / BNAIC–BeNeLearn 2023
Assessing aggressive driving behaviour.
A study of attention-based online action detection, hybrid model design, and interpretable cues in driving video.
- Authors
- J. Aechtner, A. Wilbik & M. Popa
- Published
- Springer · 2025
- DOI
- 10.1007/978-3-031-74650-5_4
The paper
Detecting actions as they unfold.
We investigated aggressive driving behaviour on the METEOR dataset using two online action-detection models, OadTR and Colar. The study evaluates their respective strengths and examines how class frequency in the training data relates to classification accuracy.
We also developed a hybrid architecture that combines categorical exemplars with future-frame prediction. To make its output more interpretable, we traced agent paths across space and time to identify salient cues behind individual predictions.